Xiaomi MiMo Benchmarks — SWE-Bench 73.4%, AIME 68.2%, Throughput & API Pricing

Last updated: July 15, 2026 · Originally published: December 18, 2025

Xiaomi MiMo performance benchmarks: 73.4% SWE-Bench Verified (V2-Flash), 68.2% AIME 2024 (MiMo-7B-RL), 1M context (V2.5-Pro), 150 tok/s throughput, $1/$3 per M tokens pricing. Frontier-competitive open-weight LLM.

4. Benchmarks and Evaluation

MiMo's benchmark performance varies meaningfully by model tier. We report the most relevant evaluations here, with comparisons to equivalent-weight-class open-weight models and select closed-source references for context.

4.1 SWE-Bench Verified

ModelSWE-Bench VerifiedParams ActiveLicense
MiMo-V2-Flash73.4%15B MoEMIT
DeepSeek-R168.3%37B denseMIT
Llama 4 Maverick (Meta)Open-weight multimodalCustom
GPT-5.5~88.7%UnknownClosed
Claude Opus 4.576.80%UnknownClosed
Claude Opus 4.8~80.9%UnknownClosed
DeepSeek-V3~39%37B MoEMIT

V2-Flash's 73.4% on SWE-Bench is the highest published result among open-weight models. For context, the top closed-source models include GPT-5.3 Codex (85.0%), Claude Opus 4.8 (~80.9%), Claude Opus 4.5 (76.80%), and Gemini 3 Flash (75.80%). This reflects the intense emphasis on code reasoning during MiMo's RL training pipeline.

4.2 AIME 2024 (Mathematics)

ModelAIME 2024Size
MiMo-7B-RL68.2%7B
DeepSeek-R1-7B65.4%7B
Qwen2.5-Math-7B62.0%7B
Math-Shepherd-7B60.1%7B

MiMo-7B-RL leads the 7B class on AIME 2024 by nearly 3 percentage points. Since AIME 2024 measures multi-step mathematical reasoning, this is a strong signal that MiMo's RL methodology produces genuinely better reasoning chains, not just memorized patterns.

4.3 Throughput and Context

Inference speed at 150 tok/s (V2-Flash on A100) places MiMo competitively with other MoE models like DeepSeek-V4 Pro and ahead of dense models in the same parameter class. The 1M context window of V2.5-Pro is among the largest available — comparable to Gemini 2.5 Pro (1M) and GPT-5.5 (1.05M).

4.4 Competitive Pricing at API Level

MiMo API pricing was permanently reduced in May 2026:

ProviderInput (per 1M tokens)Output (per 1M tokens)
MiMo-V2.5-Pro$1$3
MiMo-V2-Flash$0.50$1.50
GPT-5.5$5$30
GPT-5.4$2$12
Claude Opus 4.5$15$75
Claude Sonnet 4.6$3$15
DeepSeek-V4 Pro$0.435$0.87
DeepSeek-R1$0.55$2.19
Gemini 2.5 Pro$1.25$10
Mistral Large 3$2$6

MiMo sits in the middle of the pricing spectrum — more expensive than DeepSeek, but significantly cheaper than GPT-5.5, GPT-5.4, and Claude Opus 4.5. The MIT licensing on weights means self-hosting reduces marginal inference cost to electricity + hardware, making MiMo cost-competitive at any scale for teams with GPU infrastructure.

🧭 Research Report Chapter Navigation

Back to
📚 Research Report Hub (TL;DR + full chapter index)
Chapter 1+2+3
Architecture & Edge
Chapter 4
Benchmarks & Evaluation
Chapter 5+6
Applications & Outlook